论文笔记：MICCAI2018 Cell Detection with Star-convex Polygons

Motivation：

especially appealing for images of very crowded cells

bottom-up approach

to first classify every pixel into semantic classes (such as cell or background) and then group pixels of the same class into individual instances.

每个像素分类为语义类，然后将同一类的像素分组为单个实例。

problematic for images of very crowded cell nuclei, since only a few mis-classified pixels can cause bordering but distinct cell instances to be fused.

对于非常拥挤的细胞核图像来说，这是一个问题，因为只有少数分类错误的像素会导致接近但不同的细胞实例融合

top-down approach

first localize individual cell instances with a rough shape representation and then refine the shape in an additional step

首先使用粗略的形状表示本地化单个单元实例，然后在附加步骤中优化形状

NMS can be problematic if the objects of interest are poorly represented by their axis-aligned bounding boxes. While this can be mitigated by using rotated bounding boxes [10], it is still necessary to refine the box shape to accurately describe objects such as cell nuclei.

如果感兴趣的对象不能很好地用轴对齐的边界框表示，NMS可能会出现问题。虽然这可以通过使用旋转的边界框来缓解[10]，但仍有必要细化框的形状，以准确描述诸如细胞核之类的对象。

Method：

针对the single object class cell nuclei，predicts a shape representation

图像中每个点(i,j),分别回归star-convex polygon distances(k)和object probabilities(d).

object probabilities (d_(i,j)): 点(i,j) 的euclidean distance to the nearest background pixel。每个点一个距离

star-convex polygon distances (k_(i,j))：属于同一个物体类别的点(i,j)沿radial direction到物体边界的最远距离。每个点有k个距离

loss：

For the predicted object probabilities, binary cross-entropy loss

For the polygon distances, mean absolute error loss weighted by the ground truth object probabilities

关于star-convex polygon distances与radital detection： 

在CVPR2017 Straight to Shapes: Real-time Detection of Encoded Shapes中引入，但不适合，如下图。但是这种形状表征很适合细胞。